[dxvk] Pass present IDs to command submissions as necessary

src/dxvk/dxvk_cmdlist.cpp

@@ -50,9 +50,13 @@ namespace dxvk {
 
   VkResult DxvkCommandSubmission::submit(
           DxvkDevice*           device,
-          VkQueue               queue) {
+          VkQueue               queue,
+          uint64_t              frameId) {
     auto vk = device->vkd();
 
+    VkLatencySubmissionPresentIdNV latencyInfo = { VK_STRUCTURE_TYPE_LATENCY_SUBMISSION_PRESENT_ID_NV };
+    latencyInfo.presentID = frameId;
+
     VkSubmitInfo2 submitInfo = { VK_STRUCTURE_TYPE_SUBMIT_INFO_2 };
 
     if (!m_semaphoreWaits.empty()) {
@@ -70,6 +74,9 @@ namespace dxvk {
       submitInfo.pSignalSemaphoreInfos = m_semaphoreSignals.data();
     }
 
+    if (frameId && device->features().nvLowLatency2)
+      latencyInfo.pNext = std::exchange(submitInfo.pNext, &latencyInfo);
+
     VkResult vr = VK_SUCCESS;
 
     if (!this->isEmpty())
@@ -200,7 +207,8 @@ namespace dxvk {
   
   VkResult DxvkCommandList::submit(
     const DxvkTimelineSemaphores&       semaphores,
-          DxvkTimelineSemaphoreValues&  timelines) {
+          DxvkTimelineSemaphoreValues&  timelines,
+          uint64_t                      trackedId) {
     VkResult status = VK_SUCCESS;
 
     static const std::array<DxvkCmdBuffer, 2> SdmaCmdBuffers =
@@ -259,7 +267,7 @@ namespace dxvk {
         m_commandSubmission.signalSemaphore(semaphores.transfer,
           ++timelines.transfer, VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT);
 
-        if ((status = m_commandSubmission.submit(m_device, transfer.queueHandle)))
+        if ((status = m_commandSubmission.submit(m_device, transfer.queueHandle, trackedId)))
           return status;
 
         m_commandSubmission.waitSemaphore(semaphores.transfer,
@@ -301,7 +309,7 @@ namespace dxvk {
         ++timelines.graphics, VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT);
 
       // Finally, submit all graphics commands of the current submission
-      if ((status = m_commandSubmission.submit(m_device, graphics.queueHandle)))
+      if ((status = m_commandSubmission.submit(m_device, graphics.queueHandle, trackedId)))
         return status;
 
       // If there are WSI semaphores involved, do another submit only
@@ -311,7 +319,7 @@ namespace dxvk {
         m_commandSubmission.signalSemaphore(semaphores.graphics,
           ++timelines.graphics, VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT);
 
-        if ((status = m_commandSubmission.submit(m_device, graphics.queueHandle)))
+        if ((status = m_commandSubmission.submit(m_device, graphics.queueHandle, trackedId)))
           return status;
       }
 
@@ -321,7 +329,7 @@ namespace dxvk {
         m_commandSubmission.waitSemaphore(semaphores.graphics,
           timelines.graphics, VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT);
 
-        if (isLast && (status = m_commandSubmission.submit(m_device, transfer.queueHandle)))
+        if (isLast && (status = m_commandSubmission.submit(m_device, transfer.queueHandle, trackedId)))
           return status;
       }
     }

src/dxvk/dxvk_cmdlist.h

@@ -103,11 +103,13 @@ namespace dxvk {
      *
      * \param [in] device DXVK device
      * \param [in] queue Queue to submit to
+     * \param [in] frameId Latency frame ID
      * \returns Submission return value
      */
     VkResult submit(
             DxvkDevice*           device,
-            VkQueue               queue);
+            VkQueue               queue,
+            uint64_t              frameId);
 
     /**
      * \brief Resets object
@@ -215,11 +217,13 @@ namespace dxvk {
      *
      * \param [in] semaphores Timeline semaphore pair
      * \param [in] timelines Timeline semaphore values
+     * \param [in] frameId Latency frame ID
      * \returns Submission status
      */
     VkResult submit(
       const DxvkTimelineSemaphores&       semaphores,
-            DxvkTimelineSemaphoreValues&  timelines);
+            DxvkTimelineSemaphoreValues&  timelines,
+            uint64_t                      frameId);
     
     /**
      * \brief Stat counters

src/dxvk/dxvk_context.cpp

@@ -113,7 +113,7 @@ namespace dxvk {
   void DxvkContext::beginLatencyTracking(
     const Rc<DxvkLatencyTracker>&     tracker,
           uint64_t                    frameId) {
-    if (tracker && !m_latencyTracker) {
+    if (tracker && (!m_latencyTracker || m_latencyTracker == tracker)) {
       tracker->notifyCsRenderBegin(frameId);
 
       m_latencyTracker = tracker;

src/dxvk/dxvk_queue.cpp

@@ -131,6 +131,9 @@ namespace dxvk {
 
     std::unique_lock<dxvk::mutex> lock(m_mutex);
 
+    uint64_t trackedSubmitId = 0u;
+    uint64_t trackedPresentId = 0u;
+
     while (!m_stopped.load()) {
       m_appendCond.wait(lock, [this] {
         return m_stopped.load() || !m_submitQueue.empty();
@@ -150,10 +153,15 @@ namespace dxvk {
           m_callback(true);
 
         if (entry.submit.cmdList != nullptr) {
-          if (entry.latency.tracker)
+          if (entry.latency.tracker) {
             entry.latency.tracker->notifyQueueSubmit(entry.latency.frameId);
 
-          entry.result = entry.submit.cmdList->submit(m_semaphores, m_timelines);
+            if (!trackedSubmitId && entry.latency.frameId > trackedPresentId)
+              trackedSubmitId = entry.latency.frameId;
+          }
+
+          entry.result = entry.submit.cmdList->submit(
+            m_semaphores, m_timelines, trackedSubmitId);
           entry.timelines = m_timelines;
         } else if (entry.present.presenter != nullptr) {
           if (entry.latency.tracker)
@@ -164,6 +172,9 @@ namespace dxvk {
           if (entry.latency.tracker) {
             entry.latency.tracker->notifyQueuePresentEnd(
               entry.latency.frameId, entry.result);
+
+            trackedPresentId = entry.latency.frameId;
+            trackedSubmitId = 0u;
           }
         }